A new approach with score-level fusion for the classification of a speaker age and gender

• A system for classifying speakers according to their age and genders by combining three subsystems at score-level is proposed.
• The results of the approaches of GMM, SVM, and GMM-SV based SVM are transformed into probabilistic scores and combined.
• A performance evaluation of each subsystem is studied.
• The aGender database comprising 47-hour phone conversations with 954 speakers is used.
• Experimental results proved good in general.

In this study a new approach for classifying speakers according to their age and genders is proposed. This approach is composed of score-level fusion of seven sub-systems. In this fused system, which provides improved performance in three classification categories (age, gender and age & gender), spectral and prosodic features extracted from short-duration phone conversations are used with Gaussian Mixture Model (GMM), Support Vector Machine (SVM) and GMM supervector-based SVM classifiers. Also, by examining individual and various combinations of each system, the effect of feature types and classification methods on performance is investigated. With the proposed system, classification success rates are obtained 90.4%, 54.1%, and 53.5% in gender, age and age & gender categories respectively.

Figure optionsDownload as PowerPoint slide